The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides scripts that ingest untrusted data from command-line arguments and local files, interpolating them directly into LLM prompts without sanitization or boundary markers. * Ingestion points: conditional_workflow.sh (input argument $1) and data_processing.sh (reading raw text files via cat). * Boundary markers: None present. Content is directly concatenated into prompt strings. * Capability inventory: The skill uses the belt CLI to interact with external AI models and curl for network requests. * Sanitization: No escaping or validation is performed on the ingested content before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs the installation of external components using npx skills add, which fetches and integrates remote code from repositories like belt-sh/cli and inference-sh/skills.
[COMMAND_EXECUTION]: The skill includes shell scripts that utilize commands such as cat, mkdir, npx, and curl, which exceed the constraints specified in the frontmatter's allowed-tools field (Bash(belt *)).
[DATA_EXFILTRATION]: The monitored_workflow.sh example demonstrates a pattern for sending execution results and error data to an external, user-defined webhook via curl POST requests.

ai-automation-workflows